Unitree founder says GPT-6 shook the embodied AI industry
Unitree tested GPT-6 Astra on its own robots and called the results good without publishing numbers, while the author reviews Astra's success rates on robotic arm tasks and the gap in parameter scale.
Unitree raised ~6.1B yuan in its IPO and put nearly half toward its own robot brains. 2 months later its founder says the thing that shook the whole industry was a model from OpenAI Wang Xingxing, october 12 2026, Shanghai, China International Industry Fair: "the whole embodied intelligence industry has been shaken up after the release of GPT-6". Unitree tested it on its own robots and calls the results "good". it published no numbers most will read this as "robots by 2027". i read it as the best-known humanoid body maker saying out loud that the brain comes from somewhere else the idea: the body race is mostly run. the fight now is over who owns the brain, and how fast it can think i went through the demos, the test logs and the IPO filing numbers under the purple lamp in praga what runs under the hood: 1. THE MODEL: GPT-6 Astra. limited preview september 3 2026, paid users september 4. a ~$150 SO-101 arm drew the Golden Gate Bridge with it, checking the camera as it went. Sam Altman reposted it with "give me one too!" 2. THE GOOD NUMBER: RoboCurve put Astra on two real I2RT YAM arms with 3 cameras. red block into a bowl: 19 out of 20 (95%). the previous model they compared got 8 out of 20 3. THE BAD NUMBER: same rig, circular puzzle piece into a slot: 2 out of 20 (10%). on a Unitree Go1 at 50 Hz joint control, 250 calls gave ~5 seconds of walking because the sim had to pause between them 4. THE SIZE GAP: Astra is estimated in the trillions of parameters. the biggest VLAs in use are ~15B. π0 is 3.3B and outputs a 1-second chunk of actions in 73 ms. Astra pauses for seconds to think and is too big to run on the robot 5. THE SCOREBOARD: briefly #1 on RoboDojo, now 7th. its official result is simulation only, because real-world testing stopped after it damaged hardware. the current #1 is an LLM agent that calls π0.5. Galbot found Astra + π0.5 beat either one alone 6. THE MONEY: Unitree listed august 19 2026 at 150.80 yuan, opened at 1,100 (+629%), peaked above 400B yuan market cap. ~6.1B yuan gross, nearly half for embodied AI models. alone each number is hype or a caveat. together they say the slow genius plans, the small fast model moves, and nobody has one robot that does both yet GPT-6 ASTRA SEP 3 → $150 ARM DRAWS A BRIDGE → BLOCK 95% · PUZZLE 10% · 5 SEC OF WALKING → OCT 12 "INDUSTRY SHAKEN" → IF GPT-7 2027 → 80/80 TEST my take: he is not wrong that GPT-6 moved the field. but "good results" with no numbers, from a company that just raised billions to build its own models, is a headline, not a benchmark. and his finish line is still his own 80/80 test: ~80% of tasks in ~80% of places the robot has never seen, from voice or text alone. no public result is close to it yet the novelty is not that an LLM can steer a robot. the novelty is that the winning setup right now is two brains, a big slow one in a data center and a small fast one on the robot, and the money is betting on which one you get to own bookmark this and ask Unitree for the number behind "good" edge is open until someone shows the 80/80 test with the cameras rolling
This post quotes another post: x.com/vibelancer/status/2110337471875535334